WE-G-213CD-07: Enhancing Respiratory Motion Prediction Accuracy Using Audiovisual (AV) Biofeedback.

Abstract

Prediction of respiratory-related tumor motion is hampered by irregularities present in the patient breathing patterns. Audiovisual (AV) biofeedback reduces irregularities, thereby producing a less complex breathing pattern. The aim of this project is to improve respiratory motion prediction accuracy using an AV biofeedback system. An AV biofeedback system combined with real-time MRI was implemented in this project (4 human subjects across 5 studies (one subject had both an initial and follow-up study)). The AV biofeedback system consists of external marker positioned on the abdomen of human subjects, being tracked using an RPM system (Real-time Position Management, Varian) to guide the subject's breathing. Acquired respiratory data has been used as input for motion prediction through a dynamic multi-leaf collimator (DMLC) simulator developed by Prof. Keall. The prediction algorithm utilized was a kernel density estimation-based real-time prediction algorithm. A variety of prediction parameters were tested to determine optimum prediction performance. Prediction parameters adjusted were the delay time (DT) and training examples (TE); the parameters tested here were: DT/TE = 2500/1500, 2500/100, 1000/250, 500/250; Given that the data sampling rate was kept at 30 Hz, the resultant prediction training window lengths were 49.5, 8.25, 3.3 and 3.3seconds respectively. The mean difference between measured and predicted data for free breathing was 1.98±2.32mm; and 0.65±0.65mm for when AV biofeedback was implemented (reduction of error of 67%). The most accurate prediction results were attained using the parameters: DT/TE = 500 ms/250. This study demonstrates the improvement of respiratory motion prediction accuracy when AV biofeedback is implemented to produce a more regular breathing pattern.